Classification of Pigmented Skin Lesions using Early Fusion of Multi-Deep Features and Support Vector Machine

One of the primary organ is the skin. It protects and separates humans from various dangers. Nevertheless, the skin is susceptible to damage and can develop pigmented lesions. Manual classification of Pigmented Skin Lesions(PSL) is a major challenge due to availability and bias of physician, time consuming, invasive methods are painful and with elevated expenses. To overcome this, in this paper, two automated computer vision based approaches (using and without using dimensionality reduction) are employed by modelling multi-deep features using Support Vector Machine(SVM) for classification of PSL. The suggested system comprises of two operational stages. Initially, the performance of six Convolutional Neural Network models (VGG16, VGG19, AlexNet, Xception, ResNet50, and EfficientB3) in deep feature extraction and three types of hand-crafted features extraction methods (HOG, LBP, and SIFT) employing SVM are evaluated. The top three CNN models (AlexNet, ResNet50, and EfficientB3) based on accuracy are used to form binary and ternary early fusion hybrid models to classify the skin lesions. In the second stage, combining multiple CNN models resulted in increased dimensionality and number of features. These features are then classified using SVM once in the higher dimension and again by reducing the dimensionality by applying Principal Component Analysis(PCA). The accuracies of the combined models are then compared. It is observed that the ternary early fusion model with SVM (Alexnet+Resnet50+EfficientB3-SVM) produces the highest accuracy of 82.22% with higher number of features whereas the same model while using PCA produces the accuracy of 81.57%. Consequently, the findings indicate the suggested techniques are useful for classification of PSL.

I. Introduction

Skin, the greatest organ in the body of a person, is prone to a range of bacterial, viral, and inflammatory infections, which can result in a multitude of skin conditions and other health issues [1]. Among these, melanoma stands out as a particularly deadly kind of skin cancer, which has recently presented a serious public health risk [2]. According to recent statistics, 91,270 new cases of melanoma were observed in 2018 [1] and received treatment in the United States. The frequency and death rates of melanoma are anticipated to increase in the upcoming decades [3]. Research shows that between 2008 and 2018 there was a 53% rise in the yearly diagnostic rate of new cases of melanoma [1], [4]. Detecting melanoma early and providing suitable treatment can greatly enhance survival rates for patients [5] –[7]. Without early intervention, 99% of survivals since five decades decrease to 14% [8] –[10]. Additionally, non-melanoma skin cancer diagnoses surged by 77% between 1994 and 2014, with the most frequent sort being basal cell carcinoma, resulting in approximately 3,000 deaths annually [11]. This increasing prevalence highlights the urgent need for effective early detection and the identification of different kinds of skin cancer that can improve prognosis and stop progression [3]. Typically, melanoma is diagnosed through visual examination, which can be low in precision [12–13]. However, dermatoscopy offers a non-invasive method to capture this and dermatologists can recognize characteristics of the skin that are hidden from plain sight, thanks to high-resolution pictures of the skin [14].